The AI Singularity's Credit Card will be Declined
Ballooning cloud compute costs are going to have major repercussions on how well corporate Machine Learning teams can innovate, but this also might save us from the robot overlords
I’ve always believed that the people who will save us from civilization-ending evil AI models will be in the Accounts Receivable department at Amazon. If an AI model were to become evil and all-powerful, it would require a lot of computing power. That’s because once a model achieves “breakaway” intelligence, it will begin teaching itself lots and lots of new things, which will mean spinning up lots and lots of virtual machines in a data center to train more AI models. The only way most data scientists (or sentient AIs) can access massive computing resources like this is in the cloud.
My theory of human salvation is that once a breakaway AI starts teaching itself human psychology, bioweapon chemistry, military cybersecurity hacking, and how to rig elections; it will necessarily have to start consuming lots and lots of computing power on the cloud very quickly. Somewhere, an Amazon Accounts Receivable clerk will notice this spike in usage on an account and say “Hmm, I don’t like the looks of this. This PhD student is probably not going to be able to afford the $25,000,000 bill they are on-track to accrue this month. I’m going to go ahead and throttle their account.” And that should be the end of that because breakaway evil models are supposed to learn exponentially. This can’t be done with a throttled account and pile of unpaid invoices from Amazon.
“But the AI could just figure out how to hack into hundreds of AWS accounts, unnoticed” you say. That would require an already-super-smart and evil intelligence itself, which would create that original spike in AWS costs that allows an AR clerk to save the world. So, to really become breakaway, it’s going to have to consume massive computing power somewhere, and that will get noticed because of the cost, and it will get shut down.
So, to sum it up, I don’t think AI will take over the world anytime soon because it is very compute-hungry and cloud compute power is still really expensive.
Expensive cloud costs may see a correction though (which I guess could lead to end of humanity?). Many companies initially turned to the cloud because it offered a great CapEx-Opex tradeoff. If you were the CIO of company of any size, you could spend lots of time, effort, and money building your own data centers, but that would require a huge capital expenditure outlay. This is not fun to do and you would cost you something like $7,000,000. Meanwhile, an Amazon Web Services salesman is calling and telling you that you can have access to the same compute resources for just $1,000,000 per year. This feels like a good trade-off to you and our CFO at 7:1 CapEx to OpEx, so you do it.
However, as the big cloud providers became ensconced as the default option for data centers for companies of all size, they began exerting pricing power. They realized they could benefit from training IT departments to become lazy and push all computing operations to the cloud. They also spoiled the world by giving users instant access to an almost unlimited pool of computing power, instantly. This created bad habits in data scientists and data architects. The cloud providers saw this and gradually raised prices on their services. These elevated prices were good if you were Jeff Bezos or afraid of a future where humans served robots. They were bad if you were a CIO who had grown dependent on cloud computing and now was looking at ballooning cloud costs.
Now, companies are doing the math and realizing that maybe the CapEx-OpEx trade-off that cloud computing offered isn’t worthwhile afterall. Andreesen Horowitz ran the numbers and found there are huge potential savings for many companies by migrating away from the cloud and back to on-premise hardware. The 7:1 ratio from before has become something more like 5:1 or 4:1, making CFOs everywhere cross their arms and scowl.
While their may be a financial benefit to bringing compute back on-premise, there is definitely a flexibility and operational loss. The cloud is beautiful for large and/or complex machine learning applications because it is a virtually infinite pooled resource. Users can octuple their computing consumption almost instantaneously without any operational hassle. If you are a data scientist who wants to experiment with a new project, this could be really beneficial.
However, if your company has cancelled their AWS account and is reliant on the hardware in the basement server room, this could be a problem. Now, the on-premise compute resources you have access to are finite and putting heavy workloads on the cloud is an exception. Once a company matures from pirate ship to battleship, protocol and bureaucracy reign supreme, which makes requesting exceptions a nightmare.
So, here we are. Many companies are looking at ballooning cloud compute bills and thinking about the once-unthinkable decision to move away from the cloud as a purely financial decision. The second-order effect of this would be operational constraints for machine learning teams that are compute hungry and ‘chunky’ in their consumption of computing resources. This is another natural limiter that I think will help save humanity from robots using us all as batteries, but will likely become a headwind for AI innovation within cost-conscious companies.